Survey of multivariate calibration methods for pattern classification

Abstract

ABSTRACT Various methods for multivariate calibration like Principal Component Regression (PCR) and Partial Least Squares Regression (PLSR) are evaluated for their use in the field of pattern classification. These methods have the advantage that they can deal with high-dimensional feature spaces and multi-collinear data, since they inherently reduce the dimension of the feature space to represent it by one single dimension. Additionally, they yield very simple linear classifiers, which can be used for real-time calculation. These properties make the methods particularly useful in the field of image processing, where one often find high-dimensional spaces with linearly dependent data and usually we have tight requirements on computational complexity. Keywords: multivariate calibration, linear discrimination and pattern classification 1. INTRODUCTION In the last decades many different methods have been developed for pattern classification [1]. In terms of real-time processing pattern classification is a rarely obtained subject due to the computational effort. To speed up calculation time the algorithms must be efficient to compute like linear classifiers, which can be used for preprocessing images to find faults in textures roughly. In our investigation we evaluate several linear methods like Principal Component Regression (PCR) and Partial Least Squares Regression (PLSR) [2,8] from the field of multivariate calibration or Fisher’s Linear Discriminant Analysis (LDA) [4,5]. The main difficulty is to reduce the high dimensional feature space to a suitable subspace where the separation into two classes can be done in a simple and traceable way. The discrimination into two classes by applying a linear separator is an efficient method to preprocess a black and white image in real-time.